Clean up examples.

README.rst

@@ -79,7 +79,8 @@ install the library locally.
    python scripts/recon/admm.py
 
 
-Note (25-04-2023): for using the Pycsou models V2 has to be installed:
+Note (25-04-2023): for using reconstruction method based on Pycsou ``lensless.apgd.APGD``,
+V2 has to be installed:
 
 .. code:: bash
 

configs/defaults_recon.yaml

@@ -60,4 +60,19 @@ admm:
   mu1: 1e-6
   mu2: 1e-5
   mu3: 4e-5
-  tau: 0.0001
+  tau: 0.0001
+
+apgd:
+  # Stopping criteria
+  max_iter: 1000
+  rel_error: 1e-6
+  acceleration: True     
+  # Differentiable prior / regularization: l2, null
+  diff_penalty: null
+  diff_lambda: 0.001
+  # Proximal prior / regularization: nonneg, l1, null
+  prox_penalty: nonneg
+  prox_lambda: 0.001
+  # Lipschitz
+  lipschitz_tight: True
+  lipschitz_tol: 1.0

lensless/apgd.py

@@ -116,11 +116,13 @@ def __init__(
         dtype=np.float32,
         diff_penalty=None,
         prox_penalty=APGDPriors.NONNEG,
-        acceleration="BT",
+        acceleration=True,
         diff_lambda=0.001,
         prox_lambda=0.001,
         disp=100,
         rel_error=1e-6,
+        lipschitz_tight=True,
+        lipschitz_tol=1.0,
         **kwargs
     ):
         """
@@ -143,15 +145,20 @@ def __init__(
             Proximal functional to serve as prior / regularization term. Default
             is non-negative prior. See `Pycsou documentation <https://matthieumeo.github.io/pycsou/html/api/functionals/pycsou.func.penalty.html?highlight=penalty#module-pycsou.func.penalty>`__
             for available penalties.
-        acceleration : [None, 'BT', 'CD']
-            Which acceleration scheme should be used (None for no acceleration).
-            "BT" (Beck and Teboule) has convergence `O(1/k^2)`, while "CD"
-            (Chambolle and Dossal) has convergence `o(1/K^2)`. So "CD" should be
-            faster. but from our experience "BT" gives better results.
+        acceleration : bool, optional
+            Whether to use acceleration or not. Default is True.
         diff_lambda : float
             Weight of differentiable penalty.
         prox_lambda : float
             Weight of proximal penalty.
+        disp : int, optional
+            Display frequency. Default is 100.
+        rel_error : float, optional
+            Relative error to stop optimization. Default is 1e-6.
+        lipschitz_tight : bool, optional
+            Whether to use tight Lipschitz constant or not. Default is True.
+        lipschitz_tol : float, optional
+            Tolerance to compute Lipschitz constant. Default is 1.
         """
 
         # PSF and data are the same size / shape
@@ -168,6 +175,7 @@ def __init__(
 
         # Convolution operator
         self._H = RealFFTConvolve2D(self._psf, dtype=dtype)
+        self._H.lipschitz(tol=lipschitz_tol, tight=lipschitz_tight)
 
         # initialize solvers which will be created when data is set
         if diff_penalty is not None:
@@ -229,7 +237,7 @@ def set_data(self, data):
             stop_crit=self._stop_crit,
             track_objective=True,
             mode=pyca.solver.Mode.MANUAL,
-            acceleration=True,
+            acceleration=self._acc,
         )
 
     def reset(self):

scripts/recon/admm.py

@@ -12,7 +12,6 @@
 import time
 import pathlib as plib
 import matplotlib.pyplot as plt
-from datetime import datetime
 import numpy as np
 from lensless.io import load_data
 from lensless import ADMM
@@ -48,13 +47,7 @@ def admm(config):
         save = os.getcwd()
 
     start_time = time.time()
-    recon = ADMM(
-        psf,
-        mu1=config["admm"]["mu1"],
-        mu2=config["admm"]["mu2"],
-        mu3=config["admm"]["mu3"],
-        tau=config["admm"]["tau"],
-    )
+    recon = ADMM(psf, **config.admm)
     recon.set_data(data)
     print(f"Setup time : {time.time() - start_time} s")
 

scripts/recon/apgd_pycsou.py

@@ -29,14 +29,15 @@
 log = logging.getLogger(__name__)
 
 
-@hydra.main(version_base=None, config_path="../../configs", config_name="apgd_thumbs_up")
+@hydra.main(version_base=None, config_path="../../configs", config_name="defaults_recon")
 def apgd(
     config,
 ):
 
     psf, data = load_data(
         psf_fp=to_absolute_path(config["input"]["psf"]),
         data_fp=to_absolute_path(config["input"]["data"]),
+        dtype=config.input.dtype,
         downsample=config["preprocess"]["downsample"],
         bayer=config["preprocess"]["bayer"],
         blue_gain=config["preprocess"]["blue_gain"],
@@ -58,15 +59,7 @@ def apgd(
         save = os.getcwd()
 
     start_time = time.time()
-    recon = APGD(
-        psf=psf,
-        max_iter=config["apgd"]["max_iter"],
-        acceleration=config["apgd"]["acceleration"],
-        diff_penalty=config["apgd"]["diff_penalty"],
-        diff_lambda=config["apgd"]["diff_lambda"],
-        prox_penalty=config["apgd"]["prox_penalty"],
-        prox_lambda=config["apgd"]["prox_lambda"],
-    )
+    recon = APGD(psf=psf, **config.apgd)
     recon.set_data(data)
     print(f"Setup time : {time.time() - start_time} s")
 

scripts/recon/gradient_descent.py

@@ -13,7 +13,6 @@
 import numpy as np
 import time
 import pathlib as plib
-from datetime import datetime
 import matplotlib.pyplot as plt
 from lensless.io import load_data
 from lensless import (